Single-channel noise suppression based on a statistical source-model for speech
Niklas Harlander, Thomas Rohdenburg, Volker Hohmann · Proceedings of the International Symposium on Auditory and Audiological Research · 2007
We propose a single-channel noise suppression scheme based on a statistical source-model for speech. The scheme is adapted from Ephraim and Malah (1984) and Tchorz and Kollmeier (2003) and aims at improving short-time signal-to-noise ratio (SNR) estimates in different frequency subbands by learning and classifying auditory-model based speech signal features. First, the speech signal is transformed into so-called Amplitude-Modulation-Spectrograms (AMS) rstly described in Kollmeier and Koch (1994), which include information of both center frequencies and modulation frequencies within 32-ms analysis frames. Second, the short-time subband SNR is estimated from the AMS patterns by a neural network, which was trained based on a large speech database. A second neural net obtains nal SNR estimates from (i) the AMS-based SNR estimates by Tchorz and Kollmeier (2003), and (ii) the estimates derived from the traditional approach by Ephraim and Malah (1984). The final SNR estimates can be used to steer a Wiener filter for noise suppression. Experimental results indicate a reasonable SNR-estimation accuracy.